Inspiration
Government portals often feel fragmented and hard to navigate, especially for first-time or non-English-speaking users. We saw how even simple tasks—like filing taxes or understanding legal provisions—turn into frustrating experiences. This pushed us to build something that reduces friction and makes access more intuitive.
What it does
BrickByBrick AI is a browser-based assistant that helps users navigate government websites, starting with the Income Tax portal. It answers queries about navigation, auto-fills forms using user data, and handles general questions, including legal queries around BNS. It also supports multilingual interaction, making it more accessible.
How we built it
We combined a Chrome extension frontend with a FastAPI backend. Queries are classified using Sarvam 30B, then routed either to a Neo4j knowledge graph for navigation or a Vector RAG pipeline for legal/document queries. Databricks handles storage, embeddings, and serving, while PySpark powers the document pipeline. User profiles and chat history are stored for continuity and autofill. The auto-fill requires user's data from Databrick's databases and an LLM-Powered Agent in our Chrome Extension automates the process of filling out forms.
Challenges we ran into
Understanding messy, inconsistent government site structures was tricky. Building a reliable navigation graph required careful crawling and structuring. Handling multilingual queries while keeping responses accurate was another challenge. Also, aligning real-time context (page content) with backend reasoning wasn’t straightforward.
Accomplishments that we're proud of
We built an end-to-end system that actually works across navigation, Q&A, and autofill. The integration of knowledge graphs with RAG made responses more grounded. Multilingual support and persistent user context added real usability beyond a typical chatbot.
What we learned
We learned how combining structured (graphs) and unstructured (RAG) retrieval improves reliability. Designing for real-world messiness—bad UX, inconsistent data—is very different from clean datasets. Also, latency and orchestration become critical when multiple AI components are involved.
What's next for BricksByBricks
Expand beyond the Income Tax portal to more government services. Improve autofill accuracy with better schema mapping. Add proactive guidance (step-by-step flows instead of reactive answers). And optimize latency to make interactions feel instant.
Log in or sign up for Devpost to join the conversation.